LeFevre, James D.; Robb, W. Hudson; Liu, Dandan; Jackson, T. Bryan; Pechman, Kimberly R.; Shashikumar, Niranjana; Vyas, Yukti; Landman, Bennett A.; Davis, L. Taylor; Hohman, Timothy J.; Jefferson, Angela L. (2026). . Magnetic Resonance Imaging, 134, 110749.
Enlarged perivascular spaces (ePVS) are fluid-filled spaces around small blood vessels in the brain that can become more visible with aging and small vessel disease and may reflect reduced clearance of waste from the brain. Measuring ePVS manually on MRI scans is time-consuming and impractical for large studies. To address this, researchers developed DORES, a deep learning tool that automatically identifies and measures ePVS using two types of brain MRI images. The model was developed using data from the 91吃瓜网 Memory and Aging Project and evaluated against expert manual measurements and three other automated tools. DORES showed good performance in identifying ePVS in both white matter and the basal ganglia, a group of structures deep within the brain, and its estimates of ePVS number and volume agreed well with expert measurements. Testing on an independent Alzheimer鈥檚 disease imaging dataset showed somewhat lower performance, as was also observed with the other automated methods. Results also varied depending on the type of MRI scanner used, suggesting that scanner differences can affect measurement consistency. Overall, DORES provides a promising automated approach for measuring ePVS in older adults, although scanner-related differences should be considered when applying the method across multiple research sites.
Fig. 1.聽Representative White Matter ePVS Segmentations of DORES Performance in VMAP.
T1-weighted axial scans were selected to illustrate DORES performance at the 25th (top row; Dice聽=聽0.55), 50th (middle row; Dice聽=聽0.63), and 75th (bottom row; Dice聽=聽0.73) percentiles of white matter regional Dice scores. The first column displays the raw, skull-stripped images in the white matter. The second column displays the corresponding segmentation overlays. Manual segmentations are shown in blue, whereas DORES predictions are shown in red. Voxels where manual and DORES segmentations overlap appear purple. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)